Uthungelwano lweNeural. Iyaphi yonke lento?

Eli nqaku linamacandelo amabini:

1. Inkcazo emfutshane yolunye ulwakhiwo lwenethiwekhi yokukhangela into kwimifanekiso kunye nokwahlulwa kwemifanekiso kunye nezona khonkco ziqondakalayo kwizibonelelo kum. Ndizamile ukukhetha iinkcazo zevidiyo kwaye ngokukhethekileyo ngesiRashiya.
2. Inxalenye yesibini yinzame yokuqonda isikhokelo sophuhliso lwe-neural network architectures. Kwaye itekhnoloji esekelwe kubo.

Umzobo 1 - Ukuqonda i-neural network architectures akulula

Yonke yaqala ngokwenza izicelo zedemo ezimbini zokuhlelwa kwezinto kunye nokufunyanwa kwifowuni ye-Android:

• Idemo yangasemva, xa idatha ilungiswa kwiseva kwaye igqithiselwa kwifowuni. Ukuhlelwa kwemifanekiso yeentlobo ezintathu zeebhere: ezimdaka, ezimnyama kunye ne-teddy.
• Idemo yangaphambilixa idatha icutshungulwa kwifowuni ngokwayo. Ukufunyanwa kwezinto (ukubona into) kwiintlobo ezintathu: i-hazelnuts, amakhiwane kunye nemihla.

Kukho umahluko phakathi kwemisebenzi yokuhlelwa komfanekiso, ukufunyanwa kwezinto kumfanekiso kunye ukwahlulwa kwemifanekiso. Ke ngoko, bekukho imfuneko yokufumanisa ukuba yeyiphi i-neural network architectures efumanisa izinto kwimifanekiso kwaye yeyiphi enokwahlulahlula. Ndifumene le mizekelo ilandelayo yezakhiwo ezinezona khonkco ziqondakalayo kwizixhobo kum:

• Uluhlu lwezakhiwo ezisekelwe kwi-R-CNN (Rimimandla nge Ckwinguqu Neural NIimpawu zokusebenza): R-CNN, Fast R-CNN, Ngokukhawuleza i-R-CNN, Imaski ye-R-CNN. Ukubona into emfanekisweni, iibhokisi zokubopha zinikezelwe kusetyenziswa uThungelwano lweSindululo soMmandla (RPN) indlela. Ekuqaleni, kwasetyenziswa indlela yokukhangela ecothayo eKhethekileyo endaweni yeRPN. Emva koko imimandla emincinci ekhethiweyo iyondliwa kwigalelo lothungelwano lwe-neural oluqhelekileyo ukulungiselela ukuhlelwa. I-architecture ye-R-CNN ine-loops ecacileyo "ye" kwiindawo ezinqamlekileyo, ezifikelela kwi-2000 ehamba nge-AlexNet yenethiwekhi yangaphakathi. Iilophu ezicacileyo ze-"for" zicothisa isantya sokwenziwa komfanekiso. Inani leelophu ezicacileyo ezisebenza kuthungelwano lwe-neural lwangaphakathi liyancipha ngoguqulelo ngalunye olutsha lolwakhiwo, kwaye uninzi lolunye utshintsho lwenziwa ukunyusa isantya kunye nokutshintsha umsebenzi wokufunyanwa kwento ngokwahlulwa kwento kwiMask R-CNN.
• YOLO (You Only Look Once) yinethiwekhi yokuqala ye-neural eyaqaphela izinto ngexesha lokwenyani kwizixhobo eziphathwayo. Uphawu olwahlukileyo: ukwahlula izinto kwindlela enye (jonga nje kube kanye). Oko kukuthi, kwi-architecture ye-YOLO akukho "i-loops" ecacileyo, yingakho inethiwekhi isebenza ngokukhawuleza. Ngokomzekelo, lo mzekeliso: kwi-NumPy, xa uqhuba imisebenzi kunye ne-matrices, akukho kwakhona "i-loops" ecacileyo, ethi kwi-NumPy iphunyezwe kumanqanaba aphantsi okwakhiwa ngolwimi lweprogram ye-C. I-YOLO isebenzisa igridi yeefestile ezichazwe kwangaphambili. Ukuthintela into efanayo ukuba ichazwe ngamaxesha amaninzi, i-coefficient ye-window overlap (IoU) isetyenziswa. Iisiphambuka oKunjalo Uingonyama). Olu lwakhiwo lusebenza kuluhlu olubanzi kwaye luphezulu ukomelela: Imodeli inokuqeqeshwa kwiifoto kodwa iqhube kakuhle kwimizobo ezotywe ngesandla.
• SSD (Sigubu SMultiBox eshushu Detector) - "i-hacks" ephumelele kakhulu ye-architecture ye-YOLO isetyenzisiweyo (umzekelo, ukunyanzeliswa okungeyona i-maximum) kwaye ezintsha zongezwa ukwenza i-neural network isebenze ngokukhawuleza nangokuchanekileyo. Uphawu olwahlukileyo: Ukwahlula izinto ngendlela enye usebenzisa igridi enikiweyo yeefestile (ibhokisi ehlala ikho) kwiphiramidi yomfanekiso. Iphiramidi yomfanekiso ifakwe ngekhowudi kwi-convolution tensor ngokulandelana kweconvolution kunye nemisebenzi yokudibanisa (ngomsebenzi wokudityaniswa okuphezulu, idimension yesithuba iyahla). Ngale ndlela, zombini izinto ezinkulu nezincinci zimiselwe kwinethiwekhi enye.
• I-MobileSSD (mobileI-NetV2+ SSD) yindibaniselwano yoyilo lweneural network ezimbini. Inethiwekhi yokuqala I-MobileNetV2 isebenza ngokukhawuleza kwaye inyusa ukuchaneka kokuqaphela. I-MobileNetV2 isetyenziswa endaweni ye-VGG-16, eyayisetyenziswa ekuqaleni inqaku lokuqala. Inethiwekhi yesibini ye-SSD imisela indawo yezinto ezikumfanekiso.
• SqueezeNet -Inethiwekhi encinci kakhulu kodwa echanekileyo ye-neural. Ngokwayo, ayisombululi ingxaki yokufunyanwa kwento. Nangona kunjalo, inokusetyenziswa kwindibaniselwano yezakhiwo ezahlukeneyo. Kwaye isetyenziswe kwizixhobo eziphathwayo. Uphawu olwahlukileyo kukuba idatha iqale icinezelwe kwizihluzi ezine ze-1 × 1 ze-convolutional emva koko zandiswe zibe zine 1 × 1 kunye nezine 3 × 3 izihluzo zokuguqula. Olunye olunjalo lokuphindaphinda koxinzelelo lwedatha-ukwandiswa kubizwa ngokuba yi "Modyuli yoMlilo".
• I-DeepLab (Ukwahlulwa koMfanekiso weSemantic kunye neeNethi eziNzululwazi eziNxibelelayo) - ulwahlulo lwezinto ezikumfanekiso. Uphawu olwahlukileyo lolwakhiwo yi-convolution enwetshiweyo, egcina ukusonjululwa kwendawo. Oku kulandelwa yi-post-processing isigaba iziphumo usebenzisa imodeli probabilistic probabilistic (imeko indawo random), ekuvumela ukuba ukususa ingxolo encinane kulwahlulo kunye nokuphucula umgangatho umfanekiso ocandekileyo. Emva kwegama eloyikekayo "imodeli enokwenzeka yomzobo" ifihla isihluzo esiqhelekileyo seGaussian, esiqikelelwa ngamanqaku amahlanu.
• Wazama ukufumanisa isixhobo RefineDet (Imbumbulu enye CwangcisaINeural Network yento Iection), kodwa andizange ndiyiqonde kakhulu.
• Ndiphinde ndajonga indlela itekhnoloji “yengqalelo” esebenza ngayo: ividiyo1, ividiyo2, ividiyo3. Inqaku elahlukileyo lolwakhiwo “loluqwalaselo” lukhetho oluzenzekelayo lwemimandla yokwandiswa kwengqwalasela emfanekisweni (RoI, Rimikhosi of Iinterest) usebenzisa inethiwekhi ye-neural ebizwa ngokuba yiYunithi yokuQwalasela. Imimandla yokwandisa ingqalelo ifana neebhokisi ezibophezelayo, kodwa ngokungafaniyo nazo, aziqinanga emfanekisweni kwaye zinokuba nemida edibeneyo. Emva koko, ukusuka kwimimandla yoqwalaselo olwandisiweyo, iimpawu (iimpawu) zodwa, ezithi "zondliwa" kwiinethiwekhi ze-neural eziphindaphindiweyo kunye nezakhiwo. LSDM, GRU okanye Vanilla RNN. Uthungelwano lwe-neural oluqhelekileyo luyakwazi ukuhlalutya ubudlelwane beempawu ngokulandelelana. Uthungelwano lwe-neural oluqhelekileyo lwalusetyenziselwa ukuguqulela umbhalo kwezinye iilwimi, kwaye ngoku ukuguqulelwa imifanekiso kwisicatshulwa и umbhalo kumfanekiso.

Njengoko siphonononga olu yilo lwezakhiwo Ndaqonda ukuba andiqondi kwanto. Kwaye ayikuko ukuba inethiwekhi yam ye-neural ineengxaki kwindlela yokujonga. Ukudalwa kwazo zonke ezi zakhiwo kufana nohlobo oluthile lwe-hackathon enkulu, apho ababhali bakhuphisana kwii-hacks. Hack sisisombululo ekhawulezayo ingxaki software nzima. Oko kukuthi, akukho nxibelelwano lubonakalayo noluqondakalayo olunengqiqo phakathi kwazo zonke ezi zakhiwo. Yonke into ebadibanisayo yiseti yezona hacks ziphumeleleyo abaziboleka omnye komnye, kunye neqhelekileyo kubo bonke. umsebenzi ovaliweyo weconvolution (ukusasazwa kwempazamo, ukusasazwa komva). Hayi iindlela zokucinga! Akucaci ukuba yintoni enokwenziwa kunye nendlela yokuphucula impumelelo esele ikho.

Ngenxa yokunqongophala konxibelelwano olunengqiqo phakathi kwee-hacks, kunzima kakhulu ukuzikhumbula kunye nokusebenzisa ekusebenzeni. Olu lwazi ngamaqhekeza. Okona kulungileyo, amaxesha ambalwa anomdla kwaye angalindelekanga akhunjulwe, kodwa uninzi lwezinto eziqondwayo nezingaqondakaliyo ziyanyamalala kwinkumbulo phakathi kweentsuku ezimbalwa. Kuya kuba kuhle ukuba ngeveki ukhumbula ubuncinci igama loyilo. Kodwa kwachithwa iiyure eziliqela kwaneentsuku zokusebenza kufundwa amanqaku nokubukela iividiyo zempinda!

Umfanekiso 2 - Zoo yeNeural Networks

Uninzi lwababhali bamanqaku enzululwazi, ngokoluvo lwam lobuqu, benza konke okusemandleni ukuqinisekisa ukuba nolu lwazi luqhekekayo aluqondwa ngumfundi. Kodwa amabinzana athatha inxaxheba kwizivakalisi zemigca elishumi ezineefomula ezithatyathwe "ngaphandle emoyeni obhityileyo" sisihloko senqaku elahlukileyo (ingxaki shicilela okanye utshabalale).

Ngesi sizathu, kukho imfuneko yokucwangcisa ulwazi kusetyenziswa uthungelwano lwe-neural kwaye, ngaloo ndlela, lwandise umgangatho wokuqonda kunye nokukhumbula. Ke ngoko, esona sihloko siphambili sokuhlalutya itekhnoloji yomntu kunye noyilo lwenethiwekhi ye-neural eyenziweyo yayingumsebenzi olandelayo: fumanisa apho yonke into iya khona, kwaye ingesiso isixhobo sayo nayiphi na inethiwekhi ye-neural ethile ngokwahlukeneyo.

Iyaphi yonke lento? Iziphumo eziphambili:

• Inani lokuqaliswa kokufunda koomatshini kule minyaka mibini idlulileyo yawa kabukhali. Isizathu esinokwenzeka: "uthungelwano lwe-neural aluseyonto intsha."
• Nabani na unokudala inethiwekhi ye-neural esebenzayo ukusombulula ingxaki elula. Ukwenza oku, thatha imodeli esele ilungile kwi "zoo yemodeli" kwaye uqeqeshe umaleko wokugqibela womnatha we-neural (ukudlulisela ukufunda) kwidatha esele yenziwe ukusuka Uphendlo lweSeti yedatha kaGoogle okanye ukusuka 25 amawaka Kaggle datasets simahla ilifu Jupyter Notebook.
• Abavelisi abakhulu beenethiwekhi ze-neural baqala ukwenza "imodeli yogcino lwezilwanyana" (imodeli yezoo). Ukuzisebenzisa unokudala ngokukhawuleza isicelo sorhwebo: TF Hub yeTensorFlow, Ukufunyanwa kweMM yePyTorch, Isixhobo yeCaffe2, Chainer-modelzoo kuba Chainer kunye nezinye.
• Iinethiwekhi zeNeural ezisebenza ngaphakathi Ixesha elilungile (ngexesha lokwenyani) kwizixhobo eziphathwayo. Ukusuka kwi-10 ukuya kwi-50 izakhelo ngesekhondi.
• Ukusetyenziswa kothungelwano lwe-neural kwiifowuni (TF Lite), kwizikhangeli (TF.js) nakwi izinto zendlu (IoT, Iimva of Tiihenjisi). Ngokukodwa kwiifowuni esele zixhasa uthungelwano lwe-neural kwinqanaba le-hardware (i-neural accelerators).
• Zonke izixhobo, impahla, mhlawumbi nokutya ziya kuba nazo IP-v6 idilesi kunye nokunxibelelana omnye nomnye" - Sebastian Thrun.
• Inani leempapasho ekufundeni koomatshini sele liqalisile ukukhula ukodlula umthetho kaMoore (iphinda kabini yonke iminyaka emibini) ukusukela ngo-2015. Ngokucacileyo, sifuna iinethiwekhi ze-neural zokuhlalutya amanqaku.
• Obu buchwepheshe bulandelayo bufumana ukuthandwa:
  • I-PyTorch -Udumo lukhula ngokukhawuleza kwaye lubonakala ngathi ludlula iTensorFlow.
  • Ukukhetha okuzenzekelayo kweehyperparameters I-AutoML – udumo lukhula ngokutyibilikayo.
  • Ukuncipha kancinci kokuchaneka kunye nokwanda kwesantya sokubala: logic engaqondakaliyo, ii-algorithms ukunyusa, ukubala okungachanekanga (kuqikelelo), ubungakanani (xa izisindo zenethiwekhi ye-neural ziguqulwa zibe zii-integers kunye ne-quantized), i-neural accelerators.
  • Inguqulelo imifanekiso kwisicatshulwa и umbhalo kumfanekiso.
  • Indalo Izinto ze-3D ezivela kwividiyo, ngoku ngexesha lokwenyani.
  • Into ephambili malunga ne-DL kukuba kukho idatha eninzi, kodwa ukuqokelela nokubhala akulula. Ke ngoko, i-markup automation iyaphuhlisa (inkcazo ezenzekelayo) kuthungelwano lwe-neural usebenzisa iinethiwekhi ze-neural.
• Ngothungelwano lwe-neural, iNzululwazi yeKhompyutha ngokukhawuleza yaba isayensi yovavanyo wavuka ingxaki yokuzala.
• Imali ye-IT kunye nokuthandwa kweenethiwekhi ze-neural kwavela ngaxeshanye xa i-computing yaba lixabiso lemarike. Uqoqosho luyatshintsha ukusuka kuqoqosho lwegolide kunye nemali igolide-currency-computing. Jonga inqaku lam econophysics kunye nesizathu sokubonakala kwemali ye-IT.

Ngokuthe ngcembe kuvela entsha Indlela yokwenza inkqubo ye-ML/DL (Ukufunda ngoomatshini kunye nokuFunda okuNzulu), esekwe ekumeleni inkqubo njengeseti yeemodeli zothungelwano lwe-neural eziqeqeshiweyo.

Umzobo 3 - ML / DL njengendlela entsha yokucwangcisa inkqubo

Nangona kunjalo, ayizange ibonakale "ithiyori yenethiwekhi ye-neural", ngaphakathi onokuthi ucinge kwaye usebenze ngokucwangcisiweyo. Into ngoku ebizwa ngokuba "yithiyori" luvavanyo, i-heuristic algorithms.

Unxulumano lwam kunye nezinye izixhobo:

• Incwadana yeNzululwazi yeDatha. Ukulungiswa komfanekiso. Nabani na ofuna ukuyifumana kufuneka athumele i-imeyile (foobar167gmailcom). Ndithumela amakhonkco kumanqaku kunye neevidiyo njengoko izinto ziqokelelana.
• Ulwazi jikelele uluhlu lwezifundo kunye namanqakuendidlule kuyo nendingathanda ukudlula kuyo.
• Izifundo kunye neevidiyo zabaqalayo, apho kufuneka uqale ukufunda iinethiwekhi ze-neural. Plus incwadana "Intshayelelo yokuFunda koomatshini kunye neeNethiwekhi zeNeural ezenziweyo".
• Izixhobo eziluncedo, apho wonke umntu uya kufumana into enomdla kubo.
• Sizifumene ziluncedo kakhulu. iitshaneli zevidiyo zohlalutyo lwamanqaku enzululwazi yiNzululwazi yeDatha. Fumana, ubhalise kubo kwaye udlulisele amakhonkco kubalingane bakho kunye nam. Imizekelo:
  • Amaphepha amabini eMizuzu
  • Henry AI Labs
  • Yannic Kilcher
  • IkhowudiEmporium
  • Blogu Chengwei Zhang aka Tony607 ngemiyalelo yenyathelo ngenyathelo kunye nomthombo ovulekileyo.

Спасиalu

umthombo: www.habr.com